Training method for training machine learning algorithm, segmentation method, computer program product and segmentation device

Abstract

A training method for training a machine learning algorithm to perform image segmentation on an image of a chemical substance comprises the steps of: receiving input data of an image comprising at least a partial marker of the chemical substance, the input data identifying a shape and a location of the chemical substance; receiving a machine learning algorithm framework; training a framework using the input data to obtain a candidate machine learning algorithm for outputting predictions indicative of the shape and location of the chemical species based on the input image; and calculating a verification metric for the candidate machine learning algorithm, the verification metric being an intersection-to-union ratio (IoU) for each instance.

Description

[0001] The present invention relates to a training method for training a machine learning algorithm to perform image segmentation on images of chemical substances. The present invention further relates to a segmentation method using the trained machine learning algorithm, a computer program product and a segmentation device.

[0002] Segmentation of images representing chemical substances can be used to determine properties of the chemical substance such as shape, size, and morphology. Based on the determined properties, the preparation conditions and / or material properties of the chemical substance can be determined. Quantitative analysis of the above properties from microscopic images is usually labor-intensive. For example, if machine learning algorithms are used for segmentation, a large amount of fully annotated data is usually required.

[0003] It is an object of the present invention to improve the segmentation of images representing chemical substances.

[0004] Improving the technological properties of chemical materials or developing new ones is an important task in the chemical industry. To do this, samples need to be manufactured and analyzed. This requires a significant investment of resources in laboratory equipment as well as in analytical equipment. Image-based analysis complements other techniques. Image segmentation can be used on sample image data, allowing identification of chemical substances in the sample data, such as particles and / or different structures. Segmentation models require fully labeled images for training. In the development of new chemical materials, this is often not possible. This can be the case if not all chemical substances in the image data can be identified by their shape and position. The proposed method allows a machine learning algorithm to perform image segmentation on images of chemical substances using partially labeled images. This is beneficial in cases where not all structures in the image can be attributed to a certain label. This allows image segmentation to be performed at an early stage in the development process. This speeds up the development process.

[0005] Furthermore, fully labeled (e.g., annotated) images require a lot of work from quantitative chemists. Labeling can lead to errors, especially when structures on the image appear ambiguous. The proposed method allows for training with partially labeled images, thus reducing ambiguity and errors. This results in a more robust segmentation model.

[0006] According to a first aspect, a training method for training a machine learning algorithm to perform image segmentation on images of chemical substances is provided. The training method comprises:

[0007] receiving input data comprising an at least partially labeled image of a chemical substance, the input data identifying a shape and a position of the chemical substance on the at least partially labeled image;

[0008] receiving a machine learning algorithm framework with a corresponding set of hyperparameters;

[0009] training a machine learning algorithm framework using the input data to obtain a candidate machine learning algorithm for outputting a prediction of a shape and location indicative of a chemical substance based on an input image received as input; and

[0010] Calculating a validation metric for the candidate machine learning algorithm, the validation metric comprising an intersection over union (IoU) for each instance, wherein the IoU for each instance is a ratio of an overlap area, which is a maximum overlap area between labeled chemicals from one of the at least partially labeled images and a prediction of the candidate machine learning algorithm for a corresponding input image corresponding to one of the at least partially labeled images, and a union of the labeled chemicals from one of the at least partially labeled images and a prediction of the candidate machine learning algorithm for the corresponding input image.

[0011] The validation metric indicates the segmentation performance of the candidate machine learning algorithm, even in the case where the input data includes non-labeled (unlabeled) regions. Since the input data only needs to be partially labeled, the labeling workload can be reduced. Therefore, the number of images included in the input data can be increased, thereby advantageously improving the training of the machine learning algorithm. The trained machine learning algorithm can be used to perform segmentation on images of chemicals, thereby reducing the workload.

[0012] Chemical substances (e.g., particles) can be in the form of substances with constant chemical composition and properties. In an embodiment, chemical substances include solid particles. Other examples of chemical substances include polymers, crystals, and molecules. Polymers can be composed of many monomers joined together.

[0013] A machine learning (ML) algorithm trained according to a training method may be referred to as a "trained ML algorithm", "segmentation ML algorithm" and / or "segmentation algorithm". A trained segmentation ML algorithm allows segmenting an image received as input into different parts. The segmentation ML algorithm may be a deep learning algorithm and / or a neural network algorithm.

[0014] "Segmentation" here refers in particular to dividing an image into different regions according to the characteristics of the regions. In detail, segmentation here may refer to, for a given region or pixel of an image, identifying whether the image shows and / or belongs to a chemical substance. Segmentation may be a score (e.g., between 0 and 1) indicating the likelihood that each region or pixel of the image is part of a chemical substance.

[0015] The input data comprises at least partially labeled images, which may be considered training data. The at least partially labeled images may comprise one or more chemical markers, wherein the markers indicate the shape and position of the chemical in the image. In particular, the markers may indicate the outline of the chemical, the center of the chemical, a geometric shape close to the actual shape of the chemical, etc. For example, the marking of the chemical may be performed manually by a user. In an embodiment, the data comprising at least partially labeled images comprises at least one partially labeled image. In an embodiment, a partially labeled image may refer to an image in which the number of labeled objects is lower than the maximum number of objects in the image.

[0016] In an embodiment, the machine learning algorithm framework may refer to a machine learning model, in particular a convolutional neural network.

[0017] A machine learning algorithm framework (or model) may be defined by its corresponding set of hyperparameters, where hyperparameters may refer to parameters associated with properties of a machine learning model.

[0018] In an embodiment, the provided machine learning algorithm for performing image segmentation on images of chemical substances may be referred to as a segmentation model.

[0019] In an embodiment, a hyperparameter may refer to a parameter associated with a property of a machine learning model. The property of a machine learning model may refer to the structure of the machine learning model. This may, for example, refer to the number of layers of a neural network, in particular the structure, architecture or function of a machine learning algorithm.

[0020] Hyperparameters may be parameters or settings that can be modified to adjust the structure, architecture, and / or functionality of a machine learning algorithm framework. A set of hyperparameters may include the number of convolutional layers, dropout rate (a fraction of neurons randomly dropped), etc. as hyperparameters. Hyperparameters may be selected manually by a user or automatically, such as randomly or in a predefined order. Suggesting new hyperparameters (modifying hyperparameters) may refer to optimizing model training and may be performed by following one of the following strategies: random search, Bayesian optimization, grid search, etc.

[0021] In the training step, a machine learning algorithm framework is trained to obtain a candidate machine learning algorithm. The candidate machine learning algorithm may correspond to a machine learning algorithm framework to which certain weights (for each layer) have been assigned. The weights may be determined based on the input data. In particular, the training step for obtaining the candidate model is based on a provided set of parameters associated with the structure of the machine learning model. Therefore, during the training step, the structure of the machine learning model remains unchanged. The training step may include determining weights and / or biases between connected neurons. The general concept of training CNNs is described in "Deep Learning in Neural Networks: An Overview Technical Report" (arXiv:1404.7828v4).

[0022] The candidate machine learning algorithm may be a candidate for a trained segmentation ML algorithm. The candidate machine learning algorithm may be selected as and / or used as a trained segmentation ML algorithm. The candidate machine learning algorithm is capable of receiving input images as input. These input images may include representations of one or more chemicals. At least some of these input images are unlabeled, meaning that not all chemicals of the input images are labeled. The input images may be completely unlabeled.

[0023] The candidate ML algorithm is specifically trained so that it can perform segmentation on the input image and provide the result of the segmentation as an output. The result of the segmentation may be a prediction indicating the shape and location of the chemical substance on the input image input into the candidate ML algorithm. In particular, the prediction is a segmentation of the input image input into the candidate ML algorithm.

[0024] The candidate ML algorithm is evaluated by computing a validation metric. In particular, the validation metric indicates the performance of the candidate ML algorithm, in particular the performance of the algorithm in predicting the shape and location of chemicals in an input image input to the candidate ML algorithm.

[0025] The validation metric (validation parameter) may be or include the IoU for each instance. The IoU for each instance is slightly different from the "normal" or "standard" IoU. That is, the normal or standard IoU indicates the intersection of the labeled and predicted on the full image. The normal or standard IoU may be the number of pixels in the intersection, X, divided by the number of pixels in the union, Y.

[0026] The IoU per instance is evaluated for each labeled region separately and only the largest overlapping predicted region is used. In other words, for each ground truth instance (for each at least partially labeled image), the predicted labeled region with the maximum overlap (Omax) with the ground truth instance (G) is determined. The IoU per instance corresponds to the overlapping area (Omax, G) of the corresponding instances divided by the union area (Omax, G). The use of the proposed validation metric associated with the IoU per instance allows the use of partially labeled images. This is achieved by ignoring the predicted shape and position of chemicals in unlabeled areas of the image. Therefore, this may lead to allowing image segmentation to be performed at an early stage in the development process. This speeds up the development process.

[0027] In detail, for each at least partially labeled image, a corresponding input image can be retrieved or generated. In particular, the corresponding input image includes a representation of the same chemical substance as the corresponding at least partially labeled image, but without any labeling or without partial labeling. The corresponding input image can be used to generate the corresponding at least partially labeled image by labeling. For example, the corresponding input image can be an unlabeled version of the at least partially labeled image corresponding to it. The corresponding input image may include a representation of the same chemical substance as the corresponding at least partially labeled image, but without labeling or with fewer or more labels. The candidate ML algorithm can be applied to the corresponding input image to obtain a prediction of the shape and position of the chemical substance. The area of ​​each prediction, i.e., the area of ​​each predicted chemical substance, can be calculated. The area (size) can correspond to the number of pixels in the predicted chemical substance.

[0028] In each at least partially labeled image, the area (size) of the labeled region may be calculated, for example as a number of pixels. For an at least partially labeled image and a corresponding input image, the intersection area (overlap area) of a given labeled chemical (region) may be determined by calculating the number of pixels that intersect between the predicted chemical in the input data and the corresponding labeled chemical. The maximum overlapping area may be determined by comparing the number of pixels of each intersection. For an at least partially labeled image and a corresponding input image, the union area (common area) of a given labeled chemical (region) may be determined by calculating the union of all pixels in the input data belonging to the predicted chemical and the corresponding labeled chemical. The IoU for each instance (each image) may be determined by calculating the ratio between the intersection area and the union area of ​​the maximum overlapping area.

[0029] Compared to the normal or standard IoU, for each instance IoU, predictions in unlabeled regions (of partially labeled images) do not negatively impact the validation metric.

[0030] Thus, the validation metric indicates the segmentation performance of the candidate ML algorithm, even if the input data includes non-labeled (unlabeled) regions. Since the input data only needs to be partially labeled, the labeling workload can be reduced. Therefore, the number of images included in the input data can be increased, thereby advantageously improving the training of the ML algorithm.

[0031] According to another aspect, a training method for training a segmentation model to perform image segmentation on an image of a chemical substance is disclosed, the training method comprising:

[0032] receiving input data comprising at least partially labeled images, the at least partially labeled images comprising one or more partially labeled images of the chemical, wherein the label comprises a shape of the chemical, in particular a set of pixels associated with the chemical in the image, and a position of the chemical on the at least partially labeled images;

[0033] Receiving a machine learning model, in particular a machine learning model comprising a convolutional neural network having a corresponding set of parameters associated with the structure of the machine learning model;

[0034] Training the machine learning algorithm model using the input data to obtain a candidate segmentation model for outputting a prediction indicative of the shape and location of the chemical substance based on an input image received as input while maintaining a provided parameter set; and

[0035] Calculating a validation metric for the candidate machine learning algorithm, the validation metric comprising an intersection over union (IoU) for each instance, the IoU for each instance being a ratio of an overlap area for each instance being a maximum overlap area between labeled chemicals from one of the at least partially labeled images and a prediction of the candidate machine learning algorithm for a corresponding input image corresponding to one of the at least partially labeled images, and the union being a union of labeled chemicals from one of the at least partially labeled images and a prediction of the candidate machine learning algorithm for the corresponding input image.

[0036] In an embodiment, the training method according to claim 1 further comprises, based on the value of the calculated validation metric, in a new iteration:

[0037] providing a set of parameters associated with the structure of the segmentation model, the set of parameters being different from the set of parameters initially provided, thereby modifying the structure of the segmentation model; and

[0038] Repeating the steps of receiving the machine learning model, training the machine learning model, and calculating validation metrics for the different parameter sets associated with the structure of the machine learning model.

[0039] In an embodiment, the training method further comprises:

[0040] If the calculated validation metric is determined to be greater than or equal to a predetermined validation threshold, storing and / or outputting the current candidate machine learning algorithm as a trained machine learning algorithm for performing image segmentation; and / or

[0041] The candidate machine learning algorithm having the highest validation metric among the candidate machine learning algorithms from the multiple iterations is stored and / or outputted.

[0042] In an embodiment, a training method is disclosed herein, wherein the segmentation model is configured to acquire an image, perform a learned transformation on the image, wherein performing the learned transformation refers to segmentation, and output a list of shapes in the image; wherein the list of shapes refers to shapes identified in the image, wherein the machine learning algorithm has free parameters, particularly weights that are optimized by heuristic optimization during the training.

[0043] In an embodiment, the training method disclosed herein, wherein the machine learning model is one of the following machine learning models: U-Net or Mask-RCNN (region-based convolutional neural network).

[0044] In an embodiment, the training method disclosed herein, wherein the chemical substance is a particle made of cathode active material, nickel, cobalt and / or manganese.

[0045] In an embodiment, the training method disclosed herein, wherein the at least partially labeled image of the chemical substance is a scanning electron microscope (SEM) image.

[0046] In an embodiment, the training method disclosed herein further includes calculating an IoU score using multiple values ​​of the IoU metric, and determining a selected value of the IoU metric, the selected value of the IoU metric being the value that results in the highest IoU score among the multiple values ​​of the IoU metric.

[0047] In an embodiment, the training method disclosed herein, wherein the validation metric corresponds to an IoU score for each instance calculated using a selected value of the IoU metric.

[0048] The segmentation method according to claim 11, further comprising

[0049] Determining (S14) a technical performance parameter value of the chemical substance (3) using a performance model, wherein the performance model is parameterized based on the technical performance value and characteristics of the chemical substance, and using the determined characteristics of the chemical substance (3) as an input to the performance model;

[0050] - Provide the performance attributes of the technology.

[0051] 13. The segmentation method according to claim 12, wherein:

[0052] The performance model is a machine learning model trained using performance training data including characteristics of the chemical substance (3) and corresponding performance, and the trained performance model is configured to take the characteristics of the represented chemical substance (3) determined by image analysis as input and provide performance parameter values ​​as output.

[0053] According to an embodiment, the training method further comprises, based on the value of the calculated validation metric, (in a new iteration):

[0054] Providing different sets of hyperparameters; and

[0055] Repeat the steps of receiving a machine learning algorithm framework, training the machine learning algorithm framework, and computing validation metrics for different sets of hyperparameters.

[0056] Different sets of hyperparameters can be modified by the user or automatically updated, particularly taking into account the value of the calculated validation metric. Providing new hyperparameters (modifying hyperparameters) can refer to optimizing model training and can be performed by following one of the following strategies: random search, Bayesian optimization, grid search, etc.

[0057] In an embodiment, the training method further comprises changing a set of parameters associated with the structure of the machine learning model based on the value of the calculated validation metric. This may result in a different set of parameters associated with the structure of the machine learning model. This may be understood as changing the structure of the machine learning model. Examples may be an increase or decrease in kernel size and an increase or decrease in convolutional layers.

[0058] Using a machine learning algorithm framework with a different (new) set of hyperparameters, the following steps may be performed for different hyperparameter sets: receiving a machine learning algorithm framework, training the machine learning algorithm framework, and calculating a validation metric. The repetition of these steps may correspond to a new iteration of the training process. Performing multiple iterations of the training process may allow for improving and / or optimizing the candidate machine learning algorithm. In each iteration, the calculation of the validation metric in particular allows for evaluating the current candidate ML algorithm (of the current iteration) and deciding whether to continue training and / or how to modify the hyperparameter set.

[0059] According to a further embodiment, the training method further comprises:

[0060] If the calculated validation metric is determined to be greater than or equal to a predetermined validation threshold, storing and / or outputting the current candidate machine learning algorithm as a trained machine learning algorithm for performing image segmentation; and / or

[0061] The candidate machine learning algorithm having the highest validation metric among the candidate machine learning algorithms from the multiple iterations is stored and / or outputted.

[0062] For example, the training method may be stopped once the validation metric reaches a predetermined validation threshold. In this case, the current candidate ML algorithm (which produces such a good validation metric) may be retained as the (final) trained ML algorithm. The (final) trained ML algorithm may be stored and / or output for later use in image segmentation.

[0063] The training method may be performed until a predetermined number of training iterations is reached and / or until the validation metric stagnates. In this case, the training method may be stopped and the candidate ML algorithm that produces the highest validation metric may be retained as the (final) trained ML algorithm. The (final) trained ML algorithm may be stored and / or output for later use in image segmentation.

[0064] According to a further embodiment, the machine learning algorithm framework is a machine learning model that is configured to take an image (any input image or test image), perform a learned transformation on the image, and output a list of shapes in the image; wherein the machine learning algorithm has free parameters that are optimized during training by heuristic optimization.

[0065] In particular, "shape" here means a set of pixels describing an object (chemical substance) in an image. The free parameters may be weights. The heuristic optimization may be stochastic gradient descent. The heuristic optimization may be performed so as to be able to produce the desired output (a list of shapes) on the provided training data.

[0066] According to a further embodiment, the machine learning algorithm framework is one of the following machine learning models: U-Net or Mask-RCNN (region-based convolutional neural network).

[0067] The main difference between U-Net and Mask-RCNN is the specific way in which the learned transformation is applied to the input, but the rest of the training procedure is similar.

[0068] In the case of U-Net, the hyperparameters adjusted during training may include the dropout rate (a fraction of neurons randomly dropped), the number of convolutional layers (depth), and / or the weights of the pixel-wise loss (e.g., a higher weight for pixels separating two instances), etc. For Mask-RCNN, the hyperparameters adjusted during training include the weights of the loss function (classes, boxes, masks), etc.

[0069] According to a further embodiment, the chemical substance is particles made of cathode active material, nickel, cobalt and / or manganese.

[0070] According to a further embodiment, the at least partially marked image of the chemical substance is a scanning electron microscope (SEM) image.

[0071] According to a further embodiment, the method further comprises calculating an IoU score using a plurality of values ​​of the IoU metric, and determining a selected value of the IoU metric, the selected value of the IoU metric being a value among the plurality of values ​​of the IoU metric that results in a highest IoU score.

[0072] In particular, the outputs of U-Net and Mask-RCNN include scores (e.g., values ​​between 0 and 1 describing the likelihood that a pixel / object is a chemical). The metric used is not limited to a fixed score scale, but always measures the "best" intersection over union that can be achieved by adjusting the score threshold. This metric is the IoU metric (per-instance IoU metric). This metric defines a list of score thresholds that can be tried, such as [0.1, 0.2, ... 0.9].

[0073] An example for calculating the IoU score for each instance is as follows:

[0074] 1) For each score threshold t (i.e., for each value of the IoU metric): Calculate Sum(IoU per instance) / N_labels, where Sum(IoU per instance) is the sum of all IoUs for each instance in all N_images images, and N_labels is the number of labels. In further embodiments, the IoU of each instance less than the IoU threshold t2 may not be included in the sum. In an embodiment, for example, the IoU of each instance less than t2=0.5 is not included in the sum.

[0075] 2) The final per-instance IoU metric (i.e., the selected value of the IoU metric) selected and used is max_t IoU(t), i.e., the score threshold (the value of the IoU metric) at which the IoU of each instance has its maximum value.

[0076] Alternative IoU metrics can be considered. For example, the average of max_t IoU(t) over all images N_images can be used and stored as the selected IoU for each instance.

[0077] The selected IoU metric for each instance can be stored and used to improve the calculation and / or reliability of the validation metric. In particular, the validation metric can be calculated as a function of the IoU of each instance and the selected IoU metric for each instance. The validation metric can correspond to the IoU score of each instance calculated using the selected value of the IoU metric and can be calculated according to the method described above.

[0078] According to a further embodiment, the calculation of the IoU score comprises, for each value of the IoU metric, calculating the sum of all IoUs for each instance divided by the number of tags, wherein the IoU of each instance smaller than the value of the IoU metric is not included in the sum.

[0079] According to a further embodiment, the validation metric corresponds to an IoU score for each instance calculated using a selected value of the IoU metric.

[0080] In an embodiment, the method for training a machine learning algorithm is implemented as a software service, particularly in terms of Software as a Service (SaS) in a cloud environment, such as Amazon Web Services (AWS), Microsoft Azure, Siemens Mindsphere, etc.

[0081] One aspect includes a trained machine learning algorithm, i.e., a machine learning algorithm trained according to the training method above or below.

[0082] According to a second aspect, a segmentation method is provided for performing segmentation on data representing chemical substances using a trained machine learning algorithm trained according to the training method according to the first aspect or according to any embodiment of the first aspect. The segmentation method comprises:

[0083] receiving at least partially unlabeled data to be segmented, the at least partially unlabeled data comprising an image of the chemical;

[0084] inputting the at least partially unlabeled data into the trained machine learning algorithm; and

[0085] Label data indicating the shape and position of the chemical substance on the image of the chemical substance is output by the trained machine learning algorithm.

[0086] The segmentation method allows the use of the above-mentioned trained machine learning algorithm for segmentation purposes, in particular to identify the location (position) and shape of the chemical substance. The trained ML algorithm can receive at least partially unlabeled data to be segmented for segmentation. The at least partially unlabeled data can be a SEM image. The at least partially unlabeled data can be completely unlabeled or partially labeled, wherein the trained ML algorithm is used to find the remaining labels showing the chemical substance.

[0087] The output of the trained ML algorithm may be labelled data. In particular, the labelled data may indicate the location of the chemical substances in the image, in particular their location and shape. The labelled data may be superimposed on the image, or provided as metadata and / or data characterising the image. For each pixel, the labelled data may indicate the likelihood that the pixel belongs to a chemical substance.

[0088] According to an embodiment, the segmentation method further comprises:

[0089] Using this labeling data, image analysis is performed to determine the characteristics of the chemical represented.

[0090] It should be understood that the term "input data" in the present disclosure may refer to test data, training data, and data analyzed / evaluated by the segmentation process.

[0091] Image analysis can be performed using conventional image analysis tools. Characteristics of the chemical include size parameters (area, surface, diameter, length, perimeter, minimum raised area around the chemical, radius of the largest circle inside the chemical, etc.), shape parameters (dimensionless shape description, regardless of the absolute size of the chemical, such as compactness, sphericity, aspect ratio, shape factor, etc.), intensity-based parameters (derived from pixel intensity, such as median intensity, standard deviation of the intensity of all pixels in the region, etc.), the number of holes in the region, the number of edges in the region, the identifier of the region and / or the centroid coordinates of the region, etc.

[0092] According to a further embodiment, the segmentation method further comprises at least one of the following:

[0093] using the performance model to determine a technical performance value for the chemical substance and using the determined characteristics of the chemical substance as input to the performance model; and / or

[0094] Particle properties are predicted using a particle prediction model, and the determined characteristics of the chemical species are used as input to the particle prediction model.

[0095] According to a further embodiment,

[0096] The performance model is a machine learning model trained using performance training data including characteristics of chemical substances and corresponding performances, the trained performance model being configured to take as input the characteristics of the represented chemical substance determined by image analysis and to provide as output electrochemical performance parameters; and / or

[0097] The particle prediction model is a machine learning model trained using predictive training data, which includes process conditions of interest and corresponding particle characteristics. The trained particle prediction model is configured to take the characteristics of the represented chemical substance determined by image analysis and the process conditions of interest as input, and provide processing conditions to generate desired particles as output.

[0098] Assume that the goal of process development is to produce spherical chemicals (particles) of uniform target size. To this end, an ML model (particle prediction model) can be trained on a relevant data set (e.g., from an initial experiment, or a subset from a larger database) that takes as input the process conditions of interest (e.g., temperature, stirring speed, concentration, atmosphere, duration of process steps, etc.) and as output the average sphericity (a shape factor from the above list) and the mean and standard deviation of the particle size. Based on these models, it is possible to recommend processing conditions that should produce the desired particle characteristics.

[0099] Further, an ML model (performance model) can be trained that takes as input the features generated from the image analysis (depending on the range as well as the material composition and process conditions); and as output an electrochemical performance metric (e.g., capacity or long-term stability). From this, it can be seen if and / or which particle properties affect the performance. Some parameters that usually affect the performance are particle size distribution, particle surface area, but also the orientation of the primary crystals (uniform or disordered radially from the center of the particle).

[0100] According to another aspect, a method for estimating a property of a material or substance is presented, wherein the method comprises the following steps:

[0101] receiving images of material or substance;

[0102] segmenting the image using a first trained machine learning algorithm that outputs data indicative of the shape and location of the material or substance; and

[0103] A property of a material or substance is estimated using a second trained machine learning algorithm that receives data indicative of a shape and position of a material or substance and outputs a property parameter of the material or substance.

[0104] In an embodiment, a first machine learning algorithm is trained according to an embodiment above or below regarding a segmentation method, a second machine learning algorithm is trained according to an embodiment above or below regarding a performance model, and / or performance parameters are output as parameters of interest for a material or substance related to a chemical production process for producing the material or substance.

[0105] According to another aspect, a method for predicting a property or characteristic of a material or substance is presented, wherein the method comprises the following steps:

[0106] receiving process conditions used to manufacture materials or substances;

[0107] Using a third trained machine learning algorithm, a property or characteristic of the material or substance as a function of the process condition is predicted.

[0108] In an embodiment, a third trained machine learning algorithm further receives data indicating the shape and position of a material or substance output from the first trained machine learning algorithm, the third machine learning algorithm being trained according to the embodiments above or below regarding a particle prediction model, and / or outputting properties or characteristics of a material or substance as a function of received process conditions.

[0109] In an embodiment, the method for predicting a property or characteristic of a material or substance comprises at least one of the following steps:

[0110] Setting target properties; and

[0111] Changing process conditions;

[0112] Compare the predicted properties or characteristics of a material or substance to the target property to obtain a similarity measure.

[0113] The process conditions may be changed until the similarity measure is below a predetermined threshold. Depending on the parameterization of the target attribute, the similarity measure may be a norm, a distance, a difference, etc. The target attribute may be parameterized by a multidimensional vector. Examples of target attributes are hardness, porosity, electrical resistance, etc.

[0114] In an embodiment, the material or substance consists of solid particles.

[0115] Advantageously, instead of performing experiments, a single or a few images of a material or substance are needed to estimate its properties under various process conditions. In particular, the process conditions are taken into account by a corresponding training process of the employed machine learning algorithm. As a result, fewer resources are required in terms of laboratory capacity and transporting test samples.

[0116] These methods allow the verification of process conditions during the manufacture of the corresponding material or substance before implementation on site. Further, the methods allow the prediction of the properties of hitherto unmanufactured materials or substances. Thus, the development and research of new materials is improved.

[0117] The embodiments and features described with reference to the first aspect or the training method of an embodiment of the first aspect apply mutatis mutandis to the segmentation method of the second aspect.

[0118] According to a third aspect, a computer program product comprises instructions which, when executed by a computer, cause the computer to perform the method according to the first or second aspect or any embodiment of the first or second aspect.

[0119] The computer program product such as the computer program means may be embodied as a memory card, USB stick, CD-ROM, DVD or a file downloadable from a server in a network. For example, such a file may be provided by transmitting a file comprising the computer program product from a wireless communication network.

[0120] The embodiments and features described with reference to the method of the first and second aspects and any embodiments of the first and second aspects apply mutatis mutandis to the computer program product of the third aspect.

[0121] According to a fourth aspect, a segmentation device is provided. The segmentation device comprises:

[0122] A storage unit, the storage unit being used to store a trained machine learning algorithm trained according to the method according to the first aspect or the second aspect or any embodiment of the first aspect or the second aspect;

[0123] an input unit, the input unit being configured to receive at least partially unlabeled data to be segmented, the at least partially unlabeled data comprising an image of the chemical substance;

[0124] a processor configured to input the at least partially unlabeled data into the trained machine learning algorithm to determine labeling data indicative of a shape and a position of the chemical on the image of the chemical; and

[0125] An output unit is used to output the determined label data.

[0126] The segmentation device may be part of a computer, in particular a personal computer or an industrial computer. The storage unit storing the trained ML algorithm may be any type of temporary or permanent storage device (memory). The processor may be a central processing unit (CPU) or the like, which is configured to execute the trained ML algorithm stored therein and / or to execute the method according to the first aspect or the second aspect or any embodiment of the first aspect or the second aspect. The input unit may include a user interface to receive at least part of the unlabeled data from a user, or the input unit may be a unit that receives at least part of the unlabeled data directly from a SEM or the like. The output unit may be a user interface, such as a display, a touch screen, or the like.

[0127] The embodiments and features described with reference to the method of the first and second aspects and any embodiments of the first and second aspects apply mutatis mutandis to the segmenting device of the fourth aspect.

[0128] Further possible implementations or alternative solutions of the present invention also encompass combinations of features described above or below with respect to these embodiments (not explicitly mentioned herein). A person skilled in the art may also add separate or isolated aspects and features to the most basic form of the present invention. It should be understood that these methods may be performed as computer-implemented methods.

[0129] Further embodiments, features and advantages of the present invention will become apparent from the following description and the dependent claims, taken in conjunction with the accompanying drawings, in which:

[0130] Figure 1 A training method according to a first embodiment is shown;

[0131] Figure 2 An example of input data is shown;

[0132] Figure 3 Examples of (a) labeled chemicals, (b) predicted chemicals, and (c) overlaps and unions are shown;

[0133] Figure 4 shows a training method according to a second embodiment;

[0134] Figure 5 A segmentation method according to a first embodiment is shown;

[0135] Figure 6 An example of segmentation is shown;

[0136] Figure 7 shows a segmentation device according to an embodiment;

[0137] Figure 8 A segmentation method according to a second embodiment is shown; and

[0138] Fig. 9 shows a part of the segmentation method according to the third embodiment;

[0139] Fig.10 An example CNN is shown.

[0140] In the drawings, like reference numbers refer to identical or functionally equivalent elements unless otherwise indicated.

[0141] Figure 1 A training method according to a first embodiment is shown. The training method shown is used to train a machine learning algorithm to perform image segmentation on an image of a chemical substance. The training method can be executed by a processor of a computing system.

[0142] exist Figure 1 In step S1, input data 1 is received. An example of such input data 1 is Figure 2. The input data 1 may include multiple images 2, which are at least partially labeled images of chemical substances 3, which are particles here. In detail, the image 2 is a scanning electron microscope (SEM) image showing multiple chemical substances 3. Some, but not all, of these chemical substances 3 have been labeled by the user. The labeled chemical substances 3 have the reference number "4" and are represented in a dotted line manner. The markings indicate the position and shape of the labeled chemical substances 4. The markings provide visual indicators for pixels in the image 2 that represent chemical substances 3 and pixels that represent other things (such as background). In Figure 2 In the example of , three images 2 are shown as belonging to the input data 1, but it should be understood that the input data 1 may actually include more images 2. Further, in Figure 2 In the diagram, only some of the chemical substances 3 have reference numbers, but it should be understood that all shapes shown in the image 2 represent the same type of chemical substance 3. A type may refer to particles having the same chemical properties.

[0143] In step S2, a machine learning algorithm framework is received. The machine learning algorithm framework is characterized by a corresponding set of hyperparameters. In this example, the machine learning algorithm framework and the corresponding set of hyperparameters are selected by the user. In an alternative embodiment, they can be automatically selected according to some selection rules.

[0144] Here, the machine learning algorithm framework is the U-Net algorithm. The hyperparameters define the dropout rate (randomly dropping a fraction of neurons), the number of convolutional layers (depth), and the weights of the pixel-level loss (for example, a higher weight for pixels that separate two instances). It is also possible to define in the hyperparameters the choice of the U-Net algorithm over another type of algorithm (such as Mask-RCNN).

[0145] The received machine learning algorithm framework is a framework to be trained based on input data 1 (training data) in order to obtain a trained ML algorithm capable of performing segmentation on any input (input image or test image) showing chemical substance 3.

[0146] exist Figure 1 In step S3 of the training method, the ML algorithm framework is trained using the input data 1 to obtain a candidate ML algorithm. The candidate ML algorithm is trained so that when the algorithm receives as input an input image representing a chemical substance 3, it outputs a prediction indicating the shape and position of the chemical substance 3 in the input image. In other words, the candidate ML algorithm is able to perform segmentation on the input image representing the chemical substance 3, which corresponds to the prediction performed.

[0147] In step S4, the quality and reliability of the candidate ML algorithm is calculated as a validation metric. The validation metric may include the intersection over union (IoU) of each instance, which is the ratio of the overlap area of ​​each instance to the union.

[0148] refer to Figure 3 Explaining the IoU for each instance, the figure shows examples of (a) labeled chemicals 4, (b) predicted chemicals 5, and (c) overlap (intersection) 6 and union 7. In detail, Figure 3 (a) One of the labeled chemical substances 4 corresponding to an image 2 of input data 1. Figure 3 (b) shows the prediction 5 of the chemical substance performed by the candidate ML algorithm on the corresponding input image, which shows the same content as image 2 of the input data 1 including the label 4, but without the label 4. Therefore, Figure 3 (a) and Figure 3 (b) shows the true (actual) label 4 and the predicted label 5 of the same chemical substance 3, respectively.

[0149] Figure 3 (c) shows the actual label 4 and the predicted label 5 shown above each other. The intersection (overlap) 6 between the actual label 4 and the predicted label 5 is filled with dashed lines. The union 7 between the actual label 4 and the predicted label 5 corresponds to the intersection 6 and all areas of the substance 4 or 5 that are not in the intersection 6. These areas are represented by points. Therefore, the union 7 corresponds to Figure 3 The sum of the dashed line portion and the dashed line portion of (c). The overlap 6 and the union 7 are represented as the number of pixels included in the overlap 6 and the union 7, respectively.

[0150] The region of image 2 used to calculate the IoU for each instance is the region including the chemical 3 having the maximum overlap 6 between the label 4 and the prediction 5. The IoU for each instance of the chemical 3 having the maximum overlap 6 is calculated by dividing the overlap 6 by the union 7 (i.e., by dividing each by the number of pixels).

[0151] In other words, the IoU for each instance is the ratio of the overlap area 6 to the union 7. The overlap area 6 is the maximum overlap area between the labeled chemical 4 from one of the at least partially labeled images 2 and the prediction 5 of the candidate machine learning algorithm for the corresponding input image corresponding to one of the at least partially labeled images 2. The union 7 is the union of the labeled chemical 4 from one of the at least partially labeled images 2 and the prediction 5 of the candidate machine learning algorithm for the corresponding input image.

[0152] The higher the IoU per instance, the better the candidate ML algorithm is. By calculating the IoU per instance as a validation metric, predictions 5 in unlabeled regions of at least partially labeled images 2 do not negatively impact the validation metric. Training of the ML algorithm can thus be performed on less input data 1 (fewer images 2) and on fewer labels 4. The effort for training the ML algorithm can thus be reduced.

[0153] Figure 4 A training method according to a second embodiment is shown. Figure 4 The training methods include Figure 1 Method steps S1 to S4 are described. Figure 4 Steps S1 to S4 correspond to Figure 1 In addition to steps S1 to S4, Figure 4 Further comprising method steps S5 to S7.

[0154] In step S5, the validation metric calculated in step S4 is compared with a predetermined validation threshold. If the validation metric is determined to be equal to or greater than the predetermined validation threshold, the current candidate ML algorithm is considered to be well-performing and is retained as the final trained ML algorithm for segmentation. In this case, in step S7, the current candidate ML algorithm is stored in the database as the final trained ML algorithm.

[0155] If, in step S5, it is determined that the validation metric is less than a predetermined validation threshold, the current candidate ML algorithm is considered to be underperforming. In this case, another iteration of training is performed. In detail, in step S6, a different set of hyperparameters is provided for the machine learning algorithm framework (meaning a different machine learning algorithm framework is provided). This different set of parameters associated with the structure of the machine learning model can be generated, for example, by random search.

[0156] The different hyperparameter set is selected by the user based on the calculated validation metric. Steps S2 to S5 are then repeated using different hyperparameter sets and thus different ML algorithm frameworks. In detail, steps S2 to S6 are repeated until, in step S5, the current validation metric is equal to or above a predetermined validation threshold. In this case, as described above, at step S7, the current candidate ML algorithm is retained and stored.

[0157] As Figure 4 Alternatively or in addition to the selection process of step S5, the training method may be performed for a predetermined number of iterations (e.g., ten iterations). At the end of the ten iterations, the candidate ML algorithm with the highest validation metric is retained. The training method may also be stopped when the predetermined number of iterations is reached or the current validation metric is equal to or higher than a predetermined validation threshold (whichever occurs first).

[0158] Repeating steps S2 to S6 allows further optimization of the ML algorithm and obtaining more reliable segmentation results when using the trained ML algorithm.

[0159] The trained ML algorithm (the algorithm stored in step S7) can be used to perform segmentation on any image from the SEM. This is performed according to a segmentation method, examples of which are described in Figure 5 Shown in. Figure 5 Steps S8 to S10 of the method may be Figure 1 After steps S1 to S4 or after Figure 4 Steps S1 to S7 are then executed.

[0160] exist Figure 5 In step S8, at least part of the unlabeled data 8 to be segmented is received. Figure 6 As shown, at least part of the unlabeled data 8 includes one or more SEM images 20 of the chemical substance 3. Figure 6 In the example of , at least part of the unlabeled data 8 is completely unlabeled data, in which no chemical substances 3 are labeled. Alternatively, some but not all chemical substances 3 in the data 8 may be labeled. The purpose of the segmentation method is to segment the data 8, i.e., to identify the location and shape of the chemical substances 3 using the trained ML algorithm.

[0161] exist Figure 5 In step S9, at least part of the unlabeled data 8 is input into a trained machine learning algorithm. The trained machine learning algorithm is composed of Figure 6 Indicated by arrow 9 in FIG.

[0162] exist Figure 5 In step S10, the trained ML algorithm 9 outputs or stores label data 11 indicating the shape and position of the chemical substance 3 on the image 20 of the chemical substance 3. In detail, the label data 11 includes the prediction 5 of the trained ML algorithm 9 about the position and shape of the chemical substance 3 in the image 20, such as Figure 6 The images 20 output by the trained ML algorithm form fully labeled data 10, in which all chemicals 3 are labeled.

[0163] Instructions Figure 6 In the diagram, only some chemicals 3 and some predictions 5 have reference numerals, but it should be understood that all shown shapes of the image 20 represent chemicals 3 and all striped shapes of the image 20 represent predictions 5.

[0164] Figure 5 The segmentation method can be used Figure 7The segmentation device 12 shown in FIG. 1 is executed. In detail, the segmentation device 12 includes a storage unit 13, an input unit 14, a processor 15 and an output unit 16, all of which are connected via a connection cable 17. The storage unit 13 is a random access memory (RAM) for storing Figure 1 or Figure 4 The training method trains the trained machine learning algorithm 9.

[0165] The input unit 14 is a communication interface that receives at least partially unlabeled data 8 to be segmented directly from the SEM. The processor 15 is configured to input the at least partially unlabeled data 8 into the trained machine learning algorithm 9 to determine the label data 11 indicating the shape and position of the chemical substance 3 on the image 20 of the chemical substance 20. The processor 15 may be configured to perform Figure 1 , Figure 4 and / or Figure 5 The output unit 16 is configured to output the determined marking data 16 on a screen and / or store the marking data 16 for subsequent use.

[0166] Figure 8 A segmentation method according to the second embodiment is shown. Figure 8 The segmentation method includes steps S1 to S10 described above. Therefore, steps S1 to S10 are not described again. In addition to steps S1 to S10, Figure 8 It also includes steps S11 to S18, which will be described below.

[0167] Steps S11 to S18 are method steps performed using the segmentation results of the trained ML algorithm (particularly using the labeled data 11). In detail, in step S11, image analysis is performed using an image analysis tool to determine the characteristics of the chemical substance 3. Image analysis is performed using the labeled data 11 indicating the position and shape of the chemical substance 3. The determined characteristics of the chemical substance include size parameters (area, surface, diameter, length, circumference, minimum raised area around the chemical substance, radius of the largest circle inside the chemical substance, etc.), shape parameters (dimensionless shape description, regardless of the absolute size of the chemical substance, such as compactness, shape factor, etc.), intensity-based parameters (derived from pixel intensity, such as median density, standard deviation of the intensity of all pixels in the region, etc.), the number of holes in the region, the number of edges in the region, the identifier of the region and / or the centroid coordinates of the region, etc.

[0168] Using the features determined in step S11, two different processes can be performed, which are Figure 8 Indicated as the first example 18 and the second example 19 in Figure 8 After step S11, one or both of Examples 18 and 19 are performed.

[0169] The first example 18 relates to performance estimation using a performance model. In detail, in step S14, the features determined in step S11, the performance data (provided at step S12) and the at least partially unlabeled data 8 (provided at step S13) are input into the performance model. The performance model is a trained ML model. The performance data is data indicating the range of performance predictions and, for example, indicates desired particle characteristics, such as particle size distribution, particle surface area, orientation of primary crystals, etc. The output of the performance model is shown in step S15 and corresponds to an electrochemical performance parameter (metric), such as the capacity or long-term stability of the particle 3.

[0170] The second example 19 involves prediction of particle properties using a particle prediction model. In detail, in step S17, the features determined in step S11 or the process information or conditions (provided at step S16) are input into the particle prediction model. Optionally, the process information or conditions (provided at step S16) are also input into the previously described performance model (step S14).

[0171] The particle prediction model is a trained ML model. The process information is data about the process conditions of interest (classes of expected values, such as temperature, stirring speed, concentration, atmosphere, duration, etc.) that are planned to be applied to the particles 3, but have not yet been applied (no SEM images are available), and further includes the expected properties of the treated particles 3. The output of the particle prediction model is shown in step S18 and corresponds to the processing conditions (values ​​of temperature, stirring speed, concentration, atmosphere, duration, etc.) that achieve the desired particle characteristics. The output processing conditions allow the synthesis of particles with desired properties, thereby reducing the need for preliminary experiments or trial runs of chemical plants.

[0172] Fig. 9 FIG. 4 shows a part of the segmentation method according to the third embodiment. In detail, Fig. 9 The previously described features S6, S8 to S10 and steps S20 to S28 are shown, which may be included in Figure 8 In step S11 shown, and associated with image analysis, in order to determine the characteristics of the chemical substance 3 .

[0173] In detail, in step S20, it is determined whether the raw output of the segmentation (labeled data 11) needs to be post-processed. If post-processing is required, it is performed in step S21. The post-processing may include segmentation of the U-Net pixel scores. If post-processing is not required, or after the post-processing of step S21, in step S22, the segmented regions are obtained. In step S23, the properties of each region are quantified, which includes determining the size, shape, morphology, position, orientation, etc. of the segmented region.

[0174] exist Fig. 9 In step S24, it is determined whether the results from step S23 should be filtered. If so, in step S25, a subset of relevant areas is selected. Such relevant areas may include areas that touch the image boundary, potential mis-segmented areas, etc. After step S25, or if filtering is not required (S24), in step S26, the particles 3 are identified and the area is quantified. In step S27, a data table with the results is generated and stored. For example, the data table includes rows and columns, the rows represent each region or particle, and the columns represent each feature (including the region ID). The data generated and stored in step S27 is used to model the chemical substance 3. In step S28, this step can be performed as a substitute or supplement to step S27, providing a segmentation result and outputting it as an image for visual presentation and / or verification of the results.

[0175] Fig.10 An example of a machine learning model, particularly a convolutional neural network (CNN) for image segmentation, is shown. Depicted as 100 is an input image. In this example, the convolutional neural network includes an input layer, not shown, followed by a filter, also described as a kernel 110, which includes a kernel that slides over the input layer as a reduced matrix, thereby scanning the input image. After convolution, a feature map 120 is generated. The convolution layer convolves the input and passes its result to the next layer. In this example, the CNN further includes one or more pooling layers 120. The pooling layer reduces the dimensionality of the data by merging the output of a cluster of neurons of one layer into a single neuron of the next layer. In this example, the convolutional layer 140 and the pooling layer 150 will repeat this process, thereby further reducing the dimensionality. The CNN may further include a flattening layer 140, followed by a fully connected neural network 170, which is used to classify the fragments of the image and provide them to the output 180. The weights and biases of the CNN may be trained to obtain a candidate segmentation model. During training, the structure of the CNN remains unchanged.

[0176] Although the invention has been described in terms of preferred embodiments, it will be apparent to those skilled in the art that modifications to all embodiments may be made.

[0177] Reference numerals:

[0178] 1 Input Data

[0179] 2 At least partially labeled images

[0180] 3 Chemical substances

[0181] 4 Labelled Chemicals

[0182] 5 Predicted Chemicals

[0183] 6 Overlap

[0184] 7 Union

[0185] 8 Partially Unlabeled Data

[0186] 9 Trained Machine Learning Algorithms

[0187] 10 Fully labeled data

[0188] 11. Labeling Data

[0189] 12 Splitting equipment

[0190] 13 Storage Unit

[0191] 14 Input Unit

[0192] 15 Processor

[0193] 16 Output Units

[0194] 17 Connecting cables

[0195] 18 First Example

[0196] 19 Second Example

[0197] 20 images

[0198] S1 receives unlabeled / partially labeled image data

[0199] S2 receives a machine learning framework (ML) with hyperparameters

[0200] S3 training ML framework to obtain candidate ML algorithms

[0201] S4 Evaluating the quality of trained ML algorithms based on validation metrics

[0202] S5 compares the validation metric to the threshold

[0203] S6: Change hyperparameters / provide alternative candidate ML frameworks / algorithms

[0204] S7 stores the final / selected ML algorithm

[0205] S8 Receive (partial) labeled image data

[0206] S9 Input image data to the final / selected trained ML algorithm

[0207] S10 outputs tag data

[0208] S11 Perform image analysis / determine features

[0209] S12 provides performance data

[0210] S13 provides image data for receiving (partial) marking

[0211] S14 provides trained performance models

[0212] S15 Output performance data / parameters

[0213] S16 Receiving process information / conditions

[0214] S17 provides trained particle prediction models

[0215] S18 Output processing conditions

[0216] S20 Determine post-processing requirements

[0217] S21 performs post-processing on segmentation label data

[0218] S22 Get segmented area

[0219] S23 Quantify the properties of the segmented region

[0220] S24 Determines whether to apply the filter

[0221] S25 Select a subset of segmented regions

[0222] S26 Identify particles and quantify area

[0223] S27 Generate and store data table

[0224] S28 Output segmentation results

Claims

1. A training method for training a segmentation model (9) to perform image segmentation on an image (20) of a chemical substance (3), the training method comprising: Receiving (S1) input data (1) comprising at least partially labeled images, the at least partially labeled images comprising one or more partially labeled images (2) of the chemical substance (3), wherein the label comprises a shape of the chemical substance (3), in particular a set of pixels associated with the chemical substance in the image, and a position of the chemical substance on the at least partially labeled images (2); Receiving (S2) a machine learning model, in particular a machine learning model comprising a convolutional neural network having a corresponding set of parameters associated with the structure of the machine learning model; training (S3) the machine learning algorithm model using the input data (1) to obtain a candidate segmentation model for outputting a prediction (5) indicative of the shape and location of the chemical substance (3) based on an input image received as input, while maintaining a provided set of parameters; and Calculating (S4) a validation metric for the candidate machine learning algorithm, the validation metric comprising an intersection over union (IoU) for each instance, the IoU for each instance being a ratio of an overlap area (6) for each instance being a maximum overlap area between a labeled chemical (4) from one of the at least partially labeled images (2) and a prediction (5) of the candidate machine learning algorithm for a corresponding input image corresponding to one of the at least partially labeled images (2) to a union (7) being a union of a labeled chemical (4) from one of the at least partially labeled images (2) and a prediction (5) of the candidate machine learning algorithm for the corresponding input image.

2. The training method of claim 1 , further comprising, based on the value of the calculated validation metric, in a new iteration: providing (S6) a set of parameters associated with the structure of the segmentation model, the set of parameters being different from the set of parameters initially provided, thereby modifying the structure of the segmentation model; and Repeating the steps of receiving (S2) the machine learning model, training (S3) the machine learning model, and calculating (S4) validation metrics for the different parameter sets associated with the structure of the machine learning model.

3. The training method according to claim 1 or 2, further comprising: If the calculated validation metric is determined to be greater than or equal to a predetermined validation threshold, storing (S7) and / or outputting the current candidate machine learning algorithm as a trained machine learning algorithm (9) for performing image segmentation; and / or The candidate machine learning algorithm having the highest validation metric among the candidate machine learning algorithms from the multiple iterations is stored and / or outputted.

4. The training method according to any one of claims 1 to 3, wherein: The segmentation model is configured to acquire an image, perform a learned transformation on the image, wherein performing the learned transformation refers to segmentation, and output a list of shapes in the image; wherein the list of shapes refers to shapes identified in the image, wherein the machine learning algorithm has free parameters, specifically weights that are optimized by heuristic optimization during the training.

5. The training method according to any one of claims 1 to 4, wherein: The machine learning model is one of the following machine learning models: U-Net or Mask-RCNN (Region-based Convolutional Neural Network).

6. The training method according to any one of claims 1 to 5, wherein: The chemical substance (3) is a particle made of cathode active material, nickel, cobalt and / or manganese.

7. The training method according to any one of claims 1 to 6, wherein: The at least partially labeled image (2) of the chemical substance (3) is a scanning electron microscope (SEM) image.

8. The training method according to any one of claims 1 to 7, further comprising calculating an IoU score using multiple values ​​of the IoU metric, and determining a selected value of the IoU metric, the selected value of the IoU metric being the value among the multiple values ​​of the IoU metric that results in a highest IoU score.

9. The training method according to claim 8, wherein: This validation metric corresponds to the IoU score for each instance computed using the chosen value of this IoU metric.

10. A segmentation method for performing segmentation on data representing chemical substances (3) using a trained machine learning algorithm (9) trained according to the training method according to any one of claims 1 to 9, the segmentation method comprising: Receiving (S8) at least partially unlabeled data (8) to be segmented, the at least partially unlabeled data (8) comprising an image (20) of the chemical substance (3); Inputting (S9) the at least partially unlabeled data (8) into the trained machine learning algorithm (9); as well as The trained machine learning algorithm (9) outputs (S10) label data (11) indicating the shape and position of the chemical substance (3) on the image (20) of the chemical substance (3).

11. The segmentation method according to claim 10, further comprising: Using the labelled data (11), image analysis is performed (S11) to determine the characteristics of the chemical substance (3) represented.

12. The segmentation method according to claim 11, further comprising Determine (S14) the technical performance parameter value of the chemical substance (3) using the performance model, wherein: The performance model is parameterized based on the technical performance values ​​and characteristics of the chemical substance, and uses the determined characteristics of the chemical substance (3) as input to the performance model; - Provides performance attributes of the technology.

13. The segmentation method according to claim 12, wherein: The performance model is a machine learning model trained using performance training data including characteristics of the chemical substance (3) and corresponding performance, and the trained performance model is configured to take the characteristics of the represented chemical substance (3) determined by image analysis as input and provide performance parameter values ​​as output.

14. A computer program product comprising instructions which, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 13.

15. A segmentation device (12), comprising: A storage unit (13) for storing a trained machine learning algorithm (9) trained according to any one of claims 1 to 9; An input unit (14) for receiving at least partially unlabeled data (8) to be segmented, wherein the at least partially unlabeled data (8) comprises an image (20) of the chemical substance (3); a processor (15) configured to input the at least partially unlabeled data (8) into the trained machine learning algorithm (9) to determine label data (11) indicative of the shape and position of the chemical substance (3) on the image (20) of the chemical substance (3); as well as An output unit (16) is used to output the determined marking data (11).

Images